Abstract
Big Data evangelists often argue that algorithms make decision-making more informed and objective—a promise hotly contested by critics of these technologies. Yet, to date, most of the debate has focused on the instruments themselves, rather than on how they are used. This article addresses this lack by examining the actual practices surrounding algorithmic technologies. Specifically, drawing on multi-sited ethnographic data, I compare how algorithms are used and interpreted in two institutional contexts with markedly different characteristics: web journalism and criminal justice. I find that there are surprising similarities in how web journalists and legal professionals use algorithms in their work. In both cases, I document a gap between the intended and actual effects of algorithms—a process I analyze as “decoupling.” Second, I identify a gamut of buffering strategies used by both web journalists and legal professionals to minimize the impact of algorithms in their daily work. Those include foot-dragging, gaming, and open critique. Of course, these similarities do not exhaust the differences between the two cases, which are explored in the discussion section. I conclude with a call for further ethnographic work on algorithms in practice as an important empirical check against the dominant rhetoric of algorithmic power.
Introduction
We live in an era of data: an unprecedented amount of digital information is being collected, stored, and analyzed to predict what people do, what they think, and what they buy. Google and Facebook may be the leaders of the “Big Data revolution” (Cukier and Mayer-Schönberger, 2013), but digital technologies of quantification are also rapidly multiplying in many fields that are not directly part of the web economy. From finance (Pasquale, 2015; Poon, 2009) to healthcare (Reich, 2012), education (Espeland and Sauder, 2016; Strathern, 2000; Zeide, 2016), journalism (Anderson, 2011a), human resources (O’Neil, 2016), and criminal justice (Harcourt, 2006), algorithms and analytics are playing an increasingly important role in many expert occupations.
These developments have not gone unnoticed: a lively debate is currently taking place on the promises and limitations of algorithmic decision-making. On the one hand, Big Data evangelists emphasize the benefits of using “smart statistics” to “disrupt” or “moneyball” sectors with long histories of inefficiency and bias (Castro, 2016; Milgram, 2013). On the other hand, scholars criticize the “mythology” of Big Data (boyd and Crawford, 2012), pointing out the opacity of algorithms (Burrell, 2016; Pasquale, 2015) and delineating the discriminatory feedback loops that these “weapons of math destruction” tend to have (Barocas and Selbst, 2016; O’Neil, 2016). Many have called for increased transparency and accountability in algorithmic systems (Diakopoulos and Friedler, 2016; Pasquale, 2015).
To date, the discussion has largely focused on the instruments themselves—how algorithms are constructed and how their models operate. We know less about the practices, representations, and imaginaries of the people who rely on algorithmic technologies in their work and lives (Bucher, 2016; Gillespie, 2016; Levy, 2015). Yet there are good reasons to study the contexts of reception. Previous research in Science and Technology Studies shows that the meaning of technological artifacts is often profoundly shaped by the complex networks of people and organizations that use them (Knorr-Cetina, 1999; Orlikowski, 2007). Depending on the organizational structure and relevant social groups involved in the trajectory of a given form, innovations may either achieve rhetorical closure or fail to be implemented (Barley, 1986; Pinch and Bijker, 1984). A similar approach is now needed for algorithms. This involves exploring questions such as: How do people make sense of the recommendations provided by algorithmic tools? Do they blindly follow the algorithms’ suggestions, manipulate the instruments, or ignore them? How do algorithmic practices and representations vary depending on their context?
This article adopts a comparative approach to study the uses and meaning-making processes surrounding algorithmic tools. Specifically, I analyze how algorithms are used in two fields with different characteristics: web journalism and criminal justice. In journalism, this takes the form of real-time web analytics that provide detailed data about the behavior of online readers and make recommendations about when to promote news articles. In criminal justice, it involves a growing number of risk-assessment tools that assess the odds of recidivism of defendants. The analysis draws on multi-sited ethnographic fieldwork conducted between 2011 and 2016 in web newsrooms and criminal courts.
In spite of the many differences between web journalism and criminal justice, I document important similarities between the two cases. First, I show that in both web newsrooms and criminal courts there are discrepancies between what managers proclaim about algorithms and how workers actually use them—a process I analyze as a form of “decoupling,” drawing on organizational sociology. Second, I find that web journalists and legal professionals develop similar buffering strategies to minimize the impact of algorithms on their daily work, namely foot-dragging, gaming, and open critique. These similarities of course do not exhaust the many differences that also emerge between the two fields, which I address in the discussion section. I argue that algorithms constitute distinct kinds of symbolic resources for web journalists and legal professionals and offer tentative explanations for why this is the case. The article concludes with a call for further ethnographic work on algorithms in practice as an important empirical check against the dominant and all-encompassing rhetoric of algorithmic power.
Algorithms in expert fields
“Algorithms is a word whose time has come,” writes Mazzotti (2017). Certainly, the term algorithm now encompasses an increasingly wide—and somewhat fuzzy—set of meanings in media and academic coverage (see boyd and Crawford, 2012; Seaver, forthcoming). Here I use the word algorithm to refer to a specific type of artefact: computational procedures (which can be more or less complex) drawing on some type of digital data (“big” or not) that provide some kind of quantitative output (be it a single score or multiple metrics) through a software program. In addition, this analysis focuses on a specific subtype of algorithms: those designed to change the way in which people make decisions in their work. I will return to this point later.
This article examines the effects of algorithms, defined in this way, on what I call “expert fields.” Building on Bourdieu’s theory of fields (Bourdieu, 1993), I analyze expert fields as configurations of actors and institutions sharing a belief in the legitimacy of specific forms of knowledge as a basis for intervention in public affairs (Collins and Evans, 2007; Fourcade, 2010). Many expert fields are also professions, in the sociological sense, in that they exercise a strict monopoly on their jurisdiction and try to prevent outsiders from encroaching on their turf (Abbott, 1988). Yet expert fields differ from professions in two main ways. First, the concept of field pays closer attention to the structure of positions and position-takings within a given space of expertise (Bourdieu, 1993; Eyal and Buchholz, 2010). Second, it encompasses fuzzier occupations like journalism that do not strictly qualify as professions in the sociological sense because, unlike law or architecture, they have not established strict barriers to entry.
Discretion—or the autonomy to decide what should be done for each individual case—has long played an essential role in the daily work of experts. Experts are expected to possess some form of specialized knowledge—say, of law or medicine—and have significant autonomy in how they apply the categories of judgment deriving from this specialized knowledge. These decisions can have strong consequences on individual lives. From emergency medical procedures to criminal sentencing, experts categorize and diagnose, which can come with the power of life and death over their “customers.” Though expert knowledge has become increasingly central in contemporary societies, it is also under attack. Experts’ discretion in making decisions, which was deemed highly legitimate during the “high modernism” era of the post-World War II period (Scott, 1998), is now subject to growing critique. Experts are often blamed for the perpetuation of “broken” systems that are deemed both inefficient and discriminatory. Many expert fields that used to be protected from quantitative evaluation are now asked to comply with a growing number of metrics and standards (Espeland and Sauder, 2016; Porter, 1996; Strathern, 2000).
Algorithms reinforce this trend. Big Data evangelists often argue that accountability is not only possible but also easier than ever to achieve through automatic and computerized procedures (see for instance Ægisdottir et al., 2006; Castro, 2016; Milgram, 2013). A closer look at the discourses mobilized to justify the adoption of algorithmic techniques reveals two main arguments. First, there is an information argument: in this view, algorithms would make better decisions than individuals, simply because they have more information at their disposal, which they can compute and analyze faster and more reliably than humans. The second argument regards the purportedly objective nature of algorithmic techniques (Christin, 2016; Daston and Galison, 2007; Gillespie, 2016). In this view, algorithms would be better than humans at making decisions because they are value-neutral. In contrast to individuals, whose opinions are shaped by a variety of social factors including class, gender, race, etc., algorithms would have no politics: they would simply analyze data in the most accurate way and maximize the amount of variance explained by the model. Therefore, Big Data analytics is often prescribed as a cure for dysfunctional systems shaped by long histories of discrimination.
Unsurprisingly, these assumptions are increasingly challenged. Scholars have called for critical approaches of the “mythology” of Big Data, this “widespread belief that large data sets offer a higher form of intelligence and knowledge” (boyd and Crawford, 2012: 663). They argue that algorithmic techniques are not necessarily neutral, raising questions about biases in the training data and the disparate impact that algorithms can have on protected groups (Barocas and Selbst, 2016; O’Neil, 2016). Researchers also disagree with the promise that algorithms make experts more accountable, criticizing instead the opacity of algorithmic techniques (Burrell, 2016) and the “black boxes” that they tend to create (Pasquale, 2015).
Yet in all of this, algorithms remain somewhat decontextualized. Missing from the discussion indeed are the actual practices, applications, and uses surrounding algorithmic tools in the fields and organizations where they unfold. There is much to gain, however, from looking at such micro-level practices. Big Data is indeed only the latest attempt to change the ways in which people make decisions. As Max Weber (1978) first argued, the rationalization process started several centuries ago; the “iron cage” of modern rationality is here to stay. Scholars studying organizations know all too well that this pressure to rationalize in turn comes with many unintended consequences (DiMaggio and Powell, 1983). Organizations develop rituals and ceremonies to give the appearance of modernity without changing their practices on the ground (Meyer and Rowan, 1977). Employees and managers find ways to look compliant without changing the underlying concerns that prompted the adoption of the technology in the first place (Espeland and Sauder, 2016).
In other words, organizations are complex sites where managerial discourses and workers’ practices do not always match. Examining the local routines and meaning-making practices that take place within organizations is therefore essential to better understand the actual impact of algorithms in the social world.
Comparing web journalism and criminal justice
This article examines how experts use algorithmic tools in two fields with markedly different characteristics: web journalism and criminal justice. The two sectors match the definition of expert fields mentioned above, in the sense that it is the knowledge and experience of journalists and legal professionals that gives them legitimacy to participate in public affairs and supports their claim for autonomy in making decisions in their daily work. Yet they also differ along several important dimensions.
First, one field has a clear profit orientation, whereas the other one does not. Web journalism is structured around news organizations, which—particularly in the United States—tend to be for-profit corporations (Hallin and Mancini, 2004). In contrast, criminal justice primarily consists of public administrations. 1 In the case of web journalism, the economic motive is clear: most news organizations need to be profitable, which in online news means attracting high numbers of visitors, since most news websites—including those with subscription systems or “pay walls”—rely on online advertising as an important source of revenue. In contrast, criminal courts are public agencies whose role is primarily administrative: they enforce the law by deciding on punishment. Over the past 30 years and in parallel with the rise of mass incarceration, criminal justice budgets have been stretched to an unprecedented limit (Alexander, 2010).
Second, the two fields have different barriers to entry—or, in sociological terms, they differ in how much control they have over their jurisdictions (Abbott, 1988). In criminal justice, the practice of law is conditioned upon admission to the bar of a given state. Judges, prosecutors, and defense attorneys generally need to have a degree in order to work in criminal courts; they can be disbarred in cases of malpractice. For that reason, judges and attorneys are often described as a “pure” profession, in the sense that they have a strict control on their area of expertise. In contrast, there is no absolute criterion distinguishing journalists from non-journalists: almost anyone can start writing news articles—a fact that journalists came to resent as bloggers began to encroach on their turf (Lewis, 2012). This does not mean, however, that there are no professional norms in journalism: journalistic practice is organized around specific rules and templates, particularly in the United States, where the norm of objectivity has been a hallmark of professionalism for more than a century (Schudson, 1978).
Last but not least, the two fields differ in their orientation towards digital technologies. Journalism as an industry started moving online as early as the mid-1990s (Boczkowski, 2004). Newsrooms have always been data-rich organizations: they used to rely on telegrams and news wires to access and diffuse information; they now rely on digital content management systems and social media platforms in the daily gathering and production of information. This comes with a specific ideology regarding technology: in web newsrooms, many journalists pride themselves on being “tech-savvy.” 2 Courts are also data-rich organizations, but as most legal professionals know all too well, the practice of law has long been organized around paper archives and paper files (Vismann, 2008). Courts began using computerized case management and internal filing systems in the early 2000s, but to this day, most of the daily work done in criminal courts still involves paper-based files, which concretely means that they are many piles of paper files being carted by clerks throughout criminal courthouses. Only during the discovery phase of trials do courts heavily rely on digital tools, usually in collaboration with police departments (Bechky, 2016).
In spite of these differences, web journalism and criminal justice currently share one key feature: both fields are witnessing a multiplication of algorithmic technologies. In web journalism, this takes the form of real-time analytics software programs. In criminal courts, it involves a growing number of predictive software programs designed to evaluate the “risk” of defendants. The next section turns to these tools.
Real-time analytics in web journalism
Audience measurements are as old as media organizations themselves. Throughout the twentieth century, companies like Nielsen or Arbitron have relied on panel-based surveys to gather data about audience behavior and sell it to news organizations (Napoli, 2011). Yet when the news moved online, advertisers and marketing departments found a whole new source of information to tap into: server-based data, captured by the tracking technologies installed on a website’s servers (Turow, 2011). This in turn triggered another evolution: the emergence of audience metrics explicitly designed for editorial rather than marketing departments (Anderson, 2011a, 2011b).
A growing number of companies and programs now compile fine-grained audience data and provide visualizations for editorial use. Figure 1 provides a screenshot of Chartbeat, a real-time analytics program used by more than 80% of web publishers in the United States. Every few seconds, the Chartbeat dashboard provides up-to-date data about the number of concurrent visitors for each article, the average time spent by readers on each piece, the number of likes, shares, and tweets on Facebook and Twitter, as well as a ranking of the most popular articles. It also shows comparable information for the previous weeks and months.
Screenshot of Chartbeat. Retrieved from: https://chartbeat.com/demo/#_ on April 2, 2017.
The adoption of analytics program such as Chartbeat in web newsrooms constitutes a radical transformation of their internal dynamics. Classical ethnographies of the print era depict journalists assuming that what interested them would also interest their audience: they disregarded the letters to the editor written by unknown readers; they wrote for their colleagues, their friends, and their sources, rather than for the audience at large (Gans, 1979). In the current internet landscape, journalists cannot ignore the preferences of their readers. Web journalists and editors now routinely make predictions about which articles will be popular, expressing surprise or disappointment when the numbers do not match their expectations (Anderson, 2011a; Christin, 2014; Petre, 2015).
Taking a step back, the emergence of real-time analytics should not come as a surprise. The journalistic profession has been under attack for several decades for its elitism, political bias, and lack of diversity. Scholars have diagnosed a growing gap between the preferences of journalists and their readers (Boczkowski and Mitchelstein, 2013). Analytics companies hope to fill this gap between journalists and their audience, transforming the criteria that journalists take into account when deciding what to write and where to place it on their homepages. Take the description provided by Chartbeat: We are leaders of the real-time revolution. We partner with doers – editors, writers, marketers, developers – to deliver the data they need, when they need it. Our real-time information gives these front-line teams instantly understandable data on their users’ emergent behavior. It ends the command-and-control setup where higher-ups dictate tactical moves that the front-line should make. Instead, it empowers the front-line to be autonomous.
3
Risk-assessment tools in criminal courts
The field of criminal justice is also experiencing an algorithmic moment. Criminal justice has witnessed an exponential increase over the past 30 years: in the United States, the jail and prison population went from less than 200,000 people in 1970 to more than 2.2 million people in 2015 (Walmsley, 2016). There is growing awareness that racial bias exists at every step of the process: the current system of mass incarceration has been analyzed as a “new Jim Crow,” or system of institutionalized discrimination against African-Americans resembling the pre-Civil Rights era in the United States (Alexander, 2010; Wacquant, 2009). In many federal and state administrations, criminal justice reform is gaining traction around the argument that incarceration rates need to decrease.
This is where predictive risk-assessment tools come in. Quantitative models are far from new in criminal justice and indeed have existed for most of the twentieth century (Harcourt, 2006). Yet it is mostly over the past 20 years that courts have begun using predictive algorithms, also called risk-assessment tools, at every step of the system: more than 60 predictive instruments are currently in use in criminal courts throughout the United States (Barry-Jester et al., 2015). Based on a small number of variables about defendants, either connected to their criminal histories (previous offenses, failure to appear in court, violent offenses, etc.) or socio-demographic characteristics (age, sex, employment status, drug history, etc.), risk-assessment tools typically provide an estimate of an offender’s risk of recidivism or failure to appear when on bail, often expressed in a range of “low” to “high” risk.
Figure 2 is a screenshot of the COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) tool, a widely used risk-assessment instrument developed by a for-profit company, Northpointe Inc. The program provides a quantitative assessment of the risks of violent recidivism, general recidivism, and pretrial release, ranging from 1 (low) to 10 (high). It also shows a scale assessing the “criminogenic needs” of the defendants, for instance, their “criminal involvement,” “relationships/lifestyle,” and “personality/attitudes.”
Screenshot of COMPAS. Retrieved from: www.northpointeinc.com/files/downloads/Risk-Needs-Assessment.pdf on January 8, 2017.
Risk-assessment instruments have attracted bipartisan support among practitioners, non-profit institutions, and governmental bodies.
5
Their supporters believe these tools can to reduce overcrowding in jails by reliably identifying low-risk offenders who could be released. Take for instance the arguments developed by Anne Milgram, former Attorney General for the State of New Jersey and former Vice President of Criminal Justice at the John and Laura Arnold Foundation. In 2015, the Arnold Foundation launched a pretrial risk-assessment tool called the “Public Safety Assessment-Court” (PSA). Milgram became the public figure for the initiative and explained her views about “smart statistics” in a TED talk viewed more than 700,000 times (Milgram, 2013). Drawing a parallel with the case of baseball, where the use of data-intensive techniques transformed the game, Milgram argues that it is time to start “moneyballing justice” in order to “minimize injustice.” According to Milgram, one of the main problems of criminal courts is that they do not have enough data about defendants and inmates. Because judges and prosecutors do not have much information about these items, they rely on their “instinct” when making decisions: Judges have the best intentions when they make these decisions about risk, but they’re making them subjectively. They’re like the baseball scouts 20 years ago who were using their instinct and their experience to try to decide what risk someone poses. They’re being subjective, and we know what happens with subjective decision making, which is that we are often wrong. What we need in this space are strong data and analytics. (Milgram, 2013)
This section emphasizes how, in spite of the multiple differences that characterize web journalism and criminal justice, the two fields are currently witnessing a comparable algorithmic moment. Web analytics and risk-assessment tools match the definition of “algorithms” provided above: both are software programs drawing on digital data that perform some form of calculation and provide a quantitative output. Both sets of tools are also designed to change the ways in which web journalists and legal professionals make decisions. 6
Methods and data
To explore how web journalists and legal professionals use and make sense of algorithms in their work, I rely on multi-sited ethnographic fieldwork conducted in web newsrooms and criminal courts. Ethnographic methods are well suited for the study of work practices, especially as they relate to the construction and reception of technological artifacts (Pinch and Bijker, 1984). Ethnographers are sensitive to the discrepancies that emerge between discourses and practices in the field, or the fact that what people say does not always match what they do (Jerolmack and Khan, 2014). One potential drawback of traditional ethnographic methods is that they often sacrifice breadth for depth, focusing on a single, well-bounded site for long periods of time. Thus, scholars studying digital technologies have called for the use of “multi-sited,” “distributed,” or “networked” ethnographies (Burrell, 2009; Hine, 2007; Howard, 2002; Seaver, forthcoming) in which ethnographers follow technological artifacts and discourses as they circulate through institutional sites and epistemic networks.
This project relies on a similarly multi-sited approach but one that is not directly centered around specific technological forms or discourses. Instead, my approach is to study algorithms through their effects on workplace dynamics—a somewhat oblique strategy that I describe as “refraction ethnography,” in the sense that it focuses on how algorithms are refracted through organizational forms and work practices. I rely on grounded theory to organize my research (Charmaz, 2006). I first arrived in the field with a broad question about technology use; algorithms emerged during my fieldwork as the actors themselves repeatedly mentioned algorithmic tools during the interviews and observations. I then engaged in theoretical sampling to gather more data about this topic while trying to obtain access to organizations that differed in size and kind but still used the same set of technologies.
On the journalism side, I conducted interviews and observations between 2011 and 2015 in New York and Paris. I tried to vary the size and type of news organizations and ended up studying the use of web analytics in five newsrooms: two web magazines (one in New York, one in Paris), two “digital native” websites with a focus on news aggregation (one in New York, one in Paris), and one legacy newspaper (in Paris) that also had an online version. I conducted ethnographic observations in their newsrooms, following the daily routines of the organizations and focusing on how journalists and editors made sense of web analytics. I also interviewed 101 journalists, editors, and bloggers (45 in New York, 56 in Paris) who worked for these five organizations. I attended conferences where web journalists were speaking, read the articles they published, delved into the industry literature, and followed their exchanges on Twitter.
The research for the criminal justice segment of this project took place in two parts. I began to collect information in 2015 about the use of data technology in courts, with a specific focus on risk-assessment tools. In 2016, I conducted two weeks of observation in a midsized criminal court that I call Marcy County—a jurisdiction located in the urban area of a southern state in the United States. I complemented this fieldwork with shorter periods of observations in two other criminal courts located in large metropolitan areas on the East and West coasts. I interviewed 22 people, including probation officers, judges, defense attorneys, clerks, and technology developers. I attended conferences, workshops, “ideation” sessions, and compiled the technical literature on risk-assessment instruments. In addition, this project draws on previous ethnographic research conducted in 2005–2006 in a French criminal court (Christin, 2008). While not directly about technology, this previous research examined how judges and prosecutors made decisions about defendants, a question that informs the current project. Throughout the article, the names of individuals and organizations are changed to protect their anonymity.
Decoupling and buffering
How are algorithmic technologies used in web journalism and criminal justice? Based on the ethnographic data I collected, and in spite of the many differences between the two fields, I document several important similarities between web journalists and legal professionals.
The question of decoupling
First, when it comes to algorithms, I find that there is often a gap between what managers say their organization does and how employees in fact use these tools in their daily work. Whereas managers and executives frequently emphasize how “data-driven,” modern, and rational their organization is, the actual uses of algorithmic techniques in web newsrooms and criminal courts reveal a more complicated picture.
Take the example of TheNotebook, a web magazine based in New York (names have been changed). Like other news websites (Anderson, 2011a; Petre, 2015), TheNotebook prides itself on being data-savvy. As the editor-in-chief told me: We’re being much more conscious about data and using data. The first ten years that TheNotebook existed, we didn’t really think very much about traffic in an explicit way, people were very scared of it. So we really pushed on that. My mandate was to get more traffic… they didn’t have a clear strategy about how to do it. When I arrived people talked very differently about traffic, they saw it as threatening, debasing. That’s not the case anymore. So one of my first move was to make people more familiar with traffic and traffic numbers. Now everybody has access to Chartbeat… I also send a daily email with the highlights and successes… I also send out periodic updates. I have access to traffic numbers … They keep sending me links where I can log on and see all that stuff. In great detail (laughs)! I can’t even remember which proprietary thing they use. So, I have access to something that shows all those stats … and I don’t look at them. I just don’t bother. It’s not worth it to me. (…) I don’t go to Omniture or Chartbeat or whatever it is that they use. I don’t go there and obsessively look at the stats. I find it kind of stressful and I try not to get too wrapped up in that.
I found a comparable situation in criminal courts. During my fieldwork in Marcy County, I met with several managers and administrators who were enthusiastic about a new case management system and analytics initiative. This digital platform allowed legal professionals to upload documents online instead of filing them with the clerks; it also provided them with real-time data about their caseload and the conviction rates for their defendants. At the same time, Marcy County started relying on risk-assessment instruments to a much greater extent than before, licensing three new tools over the course of three years. As one court administrator told me, “Marcy County is shining a bright light with the analytics and the dashboards … it’s cutting edge.”
Yet, as one member of the technological team, Amelia, explained: “The technical part is not the main problem … The problem is the buy-in. It’s government, right? There is not a lot of impetus for change. We have to convince people that analytics can bring more value.” During my days of observations, I came to understand what Amelia meant. During misdemeanor and felony hearings, most judges and prosecutors did not use the analytics, dashboards, and risk-assessment tools at their disposal. Clerks and court administrators would record the proceedings on their computers, but the legal professionals involved in a given case—the defense attorneys, judges, and prosecutors—still worked on paper files. When I asked why, a court administrator told me: Yeah, here on the criminal side, it’s messy … it’s not like on the civil side. Lawyers, DAs, they negotiate, they decide on the plea … They don’t know what’s going to happen. They scribble down notes on paper sheets, they couldn’t switch to paperless! I don’t look at the numbers. There are things you can’t quantify … You can take the same case, with the same defendant, the same criminal record, the same judge, the same attorney, the same prosecutor, and get two different decisions in different courts. Or you can take the same case, with the same defendant, the same judge, etc., at a two-week interval and have completely different decision. Is that justice? I think it is.
A gamut of buffering strategies
Web journalists and legal professionals indeed develop similar strategies for minimizing the impact of algorithmic techniques in their daily work. Here I focus on three techniques that are widespread across web newsrooms and criminal courts: foot-dragging, gaming, and open critique.
A first common practice is simply to ignore the tools as much as possible, a pattern of everyday resistance analyzed by James Scott (1998) as “foot-dragging.” In web newsrooms, as noted above, journalists frequently refuse to log into Chartbeat or other analytics programs. In spite of the many emails they receive about analytics, journalists consider that it is not a central part of their job to look at traffic numbers. Consequently, they refuse to pay close attention to analytics. Similarly, in criminal courts, foot-dragging usually involves ignoring or bypassing risk scores and analytics systems. For example, in Marcy County, risk scores are systematically added to the files of the defendants by pretrial and probation officers. Yet none of the judges, prosecutors, or attorneys ever mentioned the scores during the hearings I observed. I was sometimes able to read through the files of defendants. I then realized that the risk scores sheets were usually placed towards the end of the hundred pages or so that made up the files; they were generally not annotated, in contrast to first fifty pages of the file, which frequently featured hand-written notes.
A second set of strategies has to do with “gaming,” which Espeland and Sauder define as “manipulating rules and numbers in ways that are unconnected to, or even undermine, the motivation behind them” (Espeland and Sauder, 2007: 29). In web newsrooms, a common gaming strategy with respect to analytics is the adoption of “clickbait” headlines that intentionally distort or misrepresent articles’ content in order to attract higher numbers of visitors, independently of the articles’ intrinsic appeal. Another gaming strategy used by journalists is to engage in heated negotiations with their editors to have their articles posted on top of the homepage at a popular time of the day, which mechanically increases their traffic numbers but does not change the overall performance of the website. Similarly, in criminal courts, legal professionals sometimes “select” their cases, meaning that they redirect problematic files to alternative courts, subdivisions, or attorneys in order to improve their own incarceration numbers. In addition, probation officers were found to manipulate the variables they entered in risk-assessment tools in order to obtain the score that they thought was adequate for a given defendant, a process that Hannah-Moffat et al. (2009: 405) call “criteria tinkering.”
A third kind of buffering strategy is more openly oppositional and involves open critique. Web journalists and legal professionals are often vocal in their criticism of the analytics designed for them. As John, a former prosecutor, told me when I interviewed him: They (risk-assessment instruments) are a very controversial tool. When I was a prosecutor I didn’t put much stock in it, I’d prefer to look at actual behaviors. I just didn’t know how these tests were administered, in which circumstances, with what kind of data … With these tools the output is only as good as the input. And the input is controversial. I don’t follow Chartbeat much … I try not to look at it. When you look at Chartbeat all the time, you make choices that might not be the best ones for the identity of the media … If we only cared about Chartbeat, it would be simple: we would only write about celebrities. Celebrities … it always works. But that’s not the kind of credibility we’re looking for. In the long run, when it gets too trashy, people move somewhere else.
Discussion: Algorithms as symbolic resources
So far the analysis has focused on the similarities between newsrooms and criminal courts. Yet many differences also emerge between the two cases. Within the limited space of this article, I offer a preliminary assessment of the distinct “algorithmic imaginaries” of web journalists and legal professionals and examine where they may come from.
Distinct algorithmic imaginaries
A promising way to look at how the uses and interpretations of algorithmic technologies differ between web newsrooms and criminal courts comes from the concept of the “algorithmic imaginary,” which Bucher (2016: 30) defines as “ways of thinking about what algorithms are, what they should be and how they function.” Based on my fieldwork, I argue that web journalists and legal professionals have distinct algorithmic imaginaries, in the sense that they interpret and make sense of algorithms in different ways.
Starting with the case of journalism, I realized during my interviews and observations that analytics constitute a deeply ambiguous and multifaceted object for many web journalists. Most of them do not contest that Chartbeat provide an accurate representation of their audience: they agree with the calculation methods used by real-time analytics programs. The meanings that they attach to such metrics, however, are ambivalent. Clicks first obviously correlate with economic concerns: higher traffic numbers mean more advertising revenue for the publications. Thus, clicks are seen as a signal of market success. As such, they are often criticized as an intrusion of market pressures from beyond the “wall” that is intended to protect editorial departments from commercial pressures (Gans, 1979). Yet for most journalists, traffic numbers also mean more than that. After all, many journalists want their articles to be widely read and shared. Web analytics therefore say something about the success and impact of one’s articles in the public sphere. As such, web analytics are intertwined with strong emotions, ranging from pride to shame, depending on how the article is faring in the chase for the readers’ attention.
The situation is different for legal professionals. Most of them openly contest the data and methods used to build risk-assessment tools, which they characterize as “crude” and “problematic,” and criticize the for-profit companies who designed the tools. Why should they follow the recommendations of a model built by a company that they know nothing about, using data they do not control? Many legal professionals do not see the point. For better or worse, they trust their own judgment instead. As an attorney in Marcy County told me: “It’s hard to change attitudes, because the legal profession is based on legal precedent.” In criminal justice, innovation does not come with the glitter and appeal that it has in other sectors: it is often a source of uncertainty, because by definition an innovation arrives without the vetting of precedent. Any incentive to use new tools must be balanced against the clear motivations for relying on legal tradition.
Hence, web journalists and legal professionals have different algorithmic imaginaries, which they use as symbolic resources when praising or criticizing algorithms in their daily work. These distinct meanings in turn help to understand why the algorithms under consideration are used and interpreted differently depending on their institutional context. For instance, the multifaceted meanings of clicks in web newsrooms helps make sense of why journalists do not openly resist algorithmic techniques and engage instead in interpretive work, insisting on the idea that clicks cannot be the only metric used to measure the quality of their articles. To the contrary, the stronger resistance encountered by algorithms in criminal courts makes sense given the openly dismissive opinion that most legal professionals hold about technological innovation, especially when it comes from the for-profit sector.
Exploring dimensions of difference
Why did these different imaginaries develop in web journalism and criminal justice? In this final section, I content myself with sketching several lines of analysis. As we saw in the first part of this article, web journalism and criminal justice differ in terms of their profit orientation, monopoly on their jurisdiction, and stance towards digital technologies. All of these dimensions play a role in shaping the respective imaginaries of web journalists and legal professionals.
Take the case of profit orientation. In web journalism, most staff writers know all too well that their publications need high traffic numbers to survive and that they, as employees, also need their publications to survive in order to keep their jobs. Thus, their incentives are somewhat aligned with those of editors, who are in charge of maintaining high traffic numbers (through web analytics and other means). This is not the case in criminal justice, where most judges and prosecutors—especially those who are elected, which is frequent in the United States—know and care little about budgetary constraints.
Along similar lines, the long training process and high barriers to entry in the field of law shape the professional identity of judges and prosecutors in powerful ways, making them more likely to doubt the benefits of using external tools to complement or replace their own expertise. Like John, the prosecutor mentioned earlier, legal professionals have a deep lack of trust for algorithmic tools: they explain that prefer to trust their own opinion rather than a number on a page when assessing the risk and personality of a defendant. In contrast, journalism is a particularly porous field—it is weakly professionalized, and, in Bourdieu’s terms, markedly heteronomous (Bourdieu, 1993). Consequently, journalists may have less of a strong professional identity to fall back on when asked to pay attention to web analytics and traffic numbers (Tandoc and Thomas, 2014).
Last but not least, the respective degree of immersion in digital technologies of the two fields—and more generally the different orientation of web journalists and legal professionals towards technological innovation—certainly plays a role in shaping how they make sense of algorithms. It is part of the professional identity of web journalists to care about immediacy in accessing and publishing information (think of “scoops” or “breaking news”): any technological tool that helps them achieve that goal in a faster and more reliable manner will be positively valued. Thus, journalists often enjoy the real-time aspect of web analytics. As a Parisian web editor told me, with Chartbeat, one can “see immediately how things are going […]. I need to feel what the internet users are reading, what is going on in in the world and with our readers.” In contrast, in criminal justice, the defiance against technology goes deeper than just risk-assessment instruments: it relates to the habitus of legal professionals, which is based on the rule of the precedent and the “force of law” (Bourdieu, 1987; Shklar, 1964). Technological innovation, in this view, is suspicious: it needs to become part of the tradition before it can be trusted.
These preliminary interpretations do not exhaust the many differences that emerge between the cases of web journalism and criminal justice. A more structured comparison would need to examine other important field-level and organizational features, such as the distinct role of the different actors, managers, and reformers who promote algorithms; the specific metrics that the algorithms provide; the reorganization of roles and organizational hierarchies that take place once algorithms are introduced; as well as the ways in which algorithms complement—or not—the daily work activities of experts across fields and organizations. 8
Conclusion
In spite of their many disagreements, Big Data evangelists and critical scholars have one thing in common: they tend to discuss algorithms in a somewhat decontextualized manner. This article addresses the question of context by examining the work practices that surround algorithmic technologies. Based on a multi-sited ethnographic study of how people use algorithms in two expert fields with different characteristics, I find several important similarities in what algorithms in practice look like. First, I find a gap between the intended and the actual uses of algorithms—a process I analyze as a form of decoupling. Second, similar buffering strategies can be found in web newsrooms and criminal courts to minimize the impact of algorithms on the daily work of web journalists and legal professionals. Those include foot-dragging, gaming, and open critique. In the discussion section, I turn to the differences between the two cases and argue that web journalists and legal professionals have distinct algorithmic imaginaries, which in turn are mobilized differently to comply with or resist algorithmic technologies.
To conclude, this article shows the importance of studying the practices, uses, and implementations surrounding algorithmic technologies. Intellectually, this involves establishing new exchanges between literatures that may not usually interact, such as critical data studies, the sociology of work, and organizational analysis. Politically, this is an important reminder that scholars should not accept existing discourses of algorithmic transformation at face value. Algorithms may be described as “revolutionary” (Cukier and Mayer-Schönberger, 2013), but this kind of discourse is as much prescriptive (algorithms should do all of these things) as it is descriptive. In order to push back against the rhetoric of algorithmic power, this article argues that we need to pay close attention to the actual rather than aspirational practices connected to algorithms. Further ethnographic work is needed in order to document the glitches, missteps, and losses in translation that inevitably take place when technological artifacts like algorithms are deployed into the social world.
Footnotes
Acknowledgements
The author would like to thank Morgan Ames, danah boyd, Alex Rosenblat, Anne Washington, the three anonymous reviewers, as well as the participants of the “Algorithms in Culture” conference (UC Berkeley, December 2016), the “Work, Labor, and Automation” Workshop (Data & Society Research Institute, January 2017), the Economic Sociology and Organizations Workshop (Stanford University, Department of Sociology, March 2017), and the Davis Conference on Qualitative Research (UC Davis, March 2017) for their helpful comments and feedback.
Declaration of conflicting interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
